| Message ID | 20190328152822.532-3-sibis@codeaurora.org (mailing list archive) |
|---|---|
| State | RFC, archived |
| Headers | show |
| Series | Add CPU based scaling support to Passive governor | expand |
On Thu, Mar 28, 2019 at 3:28 PM Sibi Sankar <sibis@codeaurora.org> wrote: > + > +/* The caller must call dev_pm_opp_put() after the OPP is used */ > +struct dev_pm_opp *dev_pm_opp_find_opp_of_np(struct opp_table *opp_table, > + struct device_node *opp_np) > +{ > + return _find_opp_of_np(opp_table, opp_np); > +} Hi Sibi, Though this is not the latest version, we've seen following issue: We would get lockdep warnings on this: [ 79.068957] Call trace: [ 79.071396] _find_opp_of_np+0xa0/0xa8 [ 79.075136] dev_pm_opp_find_opp_of_np+0x24/0x30 [ 79.079744] devfreq_passive_event_handler+0x304/0x51c [ 79.084872] devfreq_add_device+0x368/0x434 [ 79.089046] devm_devfreq_add_device+0x68/0xb0 [ 79.093480] mtk_cci_devfreq_probe+0x108/0x158 [ 79.097915] platform_drv_probe+0x80/0xb0 [ 79.101915] really_probe+0x1b4/0x28c [ 79.105568] driver_probe_device+0x64/0xfc [ 79.109655] __driver_attach+0x94/0xcc [ 79.113395] bus_for_each_dev+0x84/0xcc [ 79.117221] driver_attach+0x2c/0x38 [ 79.120788] bus_add_driver+0x120/0x1f4 [ 79.124614] driver_register+0x64/0xf8 [ 79.128355] __platform_driver_register+0x4c/0x58 [ 79.133049] mtk_cci_devfreq_init+0x1c/0x24 [ 79.137224] do_one_initcall+0x1c0/0x3e0 [ 79.141138] do_initcall_level+0x1f4/0x224 [ 79.145225] do_basic_setup+0x34/0x4c [ 79.148878] kernel_init_freeable+0x10c/0x194 [ 79.153225] kernel_init+0x14/0x100 [ 79.156705] ret_from_fork+0x10/0x18 [ 79.160270] irq event stamp: 238006 [ 79.163750] hardirqs last enabled at (238005): [<ffffffa71fdea0a4>] _raw_spin_unlock_irqrestore+0x40/0x84 [ 79.173391] hardirqs last disabled at (238006): [<ffffffa71f480e78>] do_debug_exception+0x70/0x198 [ 79.182337] softirqs last enabled at (237998): [<ffffffa71f48165c>] __do_softirq+0x45c/0x4a4 [ 79.190850] softirqs last disabled at (237987): [<ffffffa71f4bc0d4>] irq_exit+0xd8/0xf8 [ 79.198842] ---[ end trace 0e66a55077a0abab ]--- In _find_opp_of_np()[1], there's lockdep_assert_held(&opp_table_lock); [1] https://elixir.bootlin.com/linux/latest/source/drivers/opp/of.c#L75 But in governor passive.c#cpufreq_passive_register(), it call dev_pm_opp_find_opp_of_np() directly, so it wouldn't access opp_table_lock lock. Another similar place is in dev_pm_opp_of_add_table(), most devfreq would call this to get opp table. dev_pm_opp_of_add_table --> _opp_add_static_v2 --> _of_opp_alloc_required_opps // would goes here if opp table contains "required-opps" property. --> _find_opp_of_np cpufreq-map governor needs devfreq to have "required-opps" property. So it would also trigger above lockdep warning. The question is: Is lockdep_assert_held(&opp_table_lock); needed in above use cases? Since they don't need to modify device and opp lists. Thanks
Hi Hsin-Yi, I'll get this addressed in the next re-spin which I plan to post by end of this week. On 7/8/19 8:58 AM, Hsin-Yi Wang wrote: > On Thu, Mar 28, 2019 at 3:28 PM Sibi Sankar <sibis@codeaurora.org> wrote: > >> + >> +/* The caller must call dev_pm_opp_put() after the OPP is used */ >> +struct dev_pm_opp *dev_pm_opp_find_opp_of_np(struct opp_table *opp_table, >> + struct device_node *opp_np) >> +{ >> + return _find_opp_of_np(opp_table, opp_np); >> +} > Hi Sibi, > > Though this is not the latest version, we've seen following issue: > > We would get lockdep warnings on this: > [ 79.068957] Call trace: > [ 79.071396] _find_opp_of_np+0xa0/0xa8 > [ 79.075136] dev_pm_opp_find_opp_of_np+0x24/0x30 > [ 79.079744] devfreq_passive_event_handler+0x304/0x51c > [ 79.084872] devfreq_add_device+0x368/0x434 > [ 79.089046] devm_devfreq_add_device+0x68/0xb0 > [ 79.093480] mtk_cci_devfreq_probe+0x108/0x158 > [ 79.097915] platform_drv_probe+0x80/0xb0 > [ 79.101915] really_probe+0x1b4/0x28c > [ 79.105568] driver_probe_device+0x64/0xfc > [ 79.109655] __driver_attach+0x94/0xcc > [ 79.113395] bus_for_each_dev+0x84/0xcc > [ 79.117221] driver_attach+0x2c/0x38 > [ 79.120788] bus_add_driver+0x120/0x1f4 > [ 79.124614] driver_register+0x64/0xf8 > [ 79.128355] __platform_driver_register+0x4c/0x58 > [ 79.133049] mtk_cci_devfreq_init+0x1c/0x24 > [ 79.137224] do_one_initcall+0x1c0/0x3e0 > [ 79.141138] do_initcall_level+0x1f4/0x224 > [ 79.145225] do_basic_setup+0x34/0x4c > [ 79.148878] kernel_init_freeable+0x10c/0x194 > [ 79.153225] kernel_init+0x14/0x100 > [ 79.156705] ret_from_fork+0x10/0x18 > [ 79.160270] irq event stamp: 238006 > [ 79.163750] hardirqs last enabled at (238005): > [<ffffffa71fdea0a4>] _raw_spin_unlock_irqrestore+0x40/0x84 > [ 79.173391] hardirqs last disabled at (238006): > [<ffffffa71f480e78>] do_debug_exception+0x70/0x198 > [ 79.182337] softirqs last enabled at (237998): > [<ffffffa71f48165c>] __do_softirq+0x45c/0x4a4 > [ 79.190850] softirqs last disabled at (237987): > [<ffffffa71f4bc0d4>] irq_exit+0xd8/0xf8 > [ 79.198842] ---[ end trace 0e66a55077a0abab ]--- > > In _find_opp_of_np()[1], there's > lockdep_assert_held(&opp_table_lock); > > [1] https://elixir.bootlin.com/linux/latest/source/drivers/opp/of.c#L75 > > But in governor passive.c#cpufreq_passive_register(), it call > dev_pm_opp_find_opp_of_np() directly, so it wouldn't access > opp_table_lock lock. > > Another similar place is in dev_pm_opp_of_add_table(), most devfreq > would call this to get opp table. > dev_pm_opp_of_add_table > --> _opp_add_static_v2 > --> _of_opp_alloc_required_opps // would goes here if opp > table contains "required-opps" property. > --> _find_opp_of_np > cpufreq-map governor needs devfreq to have "required-opps" property. > So it would also trigger above lockdep warning. > > > The question is: Is lockdep_assert_held(&opp_table_lock); needed in > above use cases? Since they don't need to modify device and opp lists. > > Thanks > > >
diff --git a/drivers/opp/of.c b/drivers/opp/of.c index 539f3d013a59..d9d8875eca05 100644 --- a/drivers/opp/of.c +++ b/drivers/opp/of.c @@ -98,8 +98,8 @@ static struct dev_pm_opp *_find_opp_of_np(struct opp_table *opp_table, return NULL; } -static struct device_node *of_parse_required_opp(struct device_node *np, - int index) +struct device_node *of_parse_required_opp(struct device_node *np, + int index) { struct device_node *required_np; @@ -111,6 +111,15 @@ static struct device_node *of_parse_required_opp(struct device_node *np, return required_np; } +EXPORT_SYMBOL_GPL(of_parse_required_opp); + +/* The caller must call dev_pm_opp_put() after the OPP is used */ +struct dev_pm_opp *dev_pm_opp_find_opp_of_np(struct opp_table *opp_table, + struct device_node *opp_np) +{ + return _find_opp_of_np(opp_table, opp_np); +} +EXPORT_SYMBOL_GPL(dev_pm_opp_find_opp_of_np); /* The caller must call dev_pm_opp_put_opp_table() after the table is used */ static struct opp_table *_find_table_of_opp_np(struct device_node *opp_np) diff --git a/include/linux/pm_opp.h b/include/linux/pm_opp.h index 82ff8e2e1ff7..d7cb0e65c4f0 100644 --- a/include/linux/pm_opp.h +++ b/include/linux/pm_opp.h @@ -359,8 +359,11 @@ void dev_pm_opp_of_cpumask_remove_table(const struct cpumask *cpumask); int dev_pm_opp_of_get_sharing_cpus(struct device *cpu_dev, struct cpumask *cpumask); struct device_node *dev_pm_opp_of_get_opp_desc_node(struct device *dev); struct device_node *dev_pm_opp_get_of_node(struct dev_pm_opp *opp); +struct device_node *of_parse_required_opp(struct device_node *np, int index); int of_get_required_opp_performance_state(struct device_node *np, int index); void dev_pm_opp_of_register_em(struct cpumask *cpus); +struct dev_pm_opp *dev_pm_opp_find_opp_of_np(struct opp_table *opp_table, + struct device_node *opp_np); #else static inline int dev_pm_opp_of_add_table(struct device *dev) { @@ -390,6 +393,11 @@ static inline int dev_pm_opp_of_get_sharing_cpus(struct device *cpu_dev, struct return -ENOTSUPP; } +static inline struct dev_pm_opp *dev_pm_opp_find_opp_of_np(struct opp_table *opp_table, struct device_node *opp_np) +{ + return ERR_PTR(-ENOTSUPP); +} + static inline struct device_node *dev_pm_opp_of_get_opp_desc_node(struct device *dev) { return NULL; @@ -408,6 +416,11 @@ static inline int of_get_required_opp_performance_state(struct device_node *np, { return -ENOTSUPP; } + +static inline struct device_node *of_parse_required_opp(struct device_node *np, int index) +{ + return NULL; +} #endif #endif /* __LINUX_OPP_H__ */
Export 'dev_pm_opp_find_opp_of_np' and 'of_parse_required_nodes' as it will be used by passive governor to parse and auto-populate mapping specified using the required-opps property. Signed-off-by: Sibi Sankar <sibis@codeaurora.org> --- drivers/opp/of.c | 13 +++++++++++-- include/linux/pm_opp.h | 13 +++++++++++++ 2 files changed, 24 insertions(+), 2 deletions(-)